Title:Factorised Representations of Join Queries: Tight Bounds and a New Dichotomy

Abstract:A common theme in factorised databases and knowledge compilation is the representation of solution sets in a useful yet succinct data structure. In this paper, we study the representation of the result of join queries (or, equivalently, the set of homomorphisms between two relational structures). We focus on the very general format of $\{\cup, \times\}$-circuits -- also known as d-representations or DNNF circuits -- and aim to find the limits of this approach.
In prior work, it has been shown that there always exists a $\{\cup, \times\}$-circuits-circuit of size $N^{O(subw)}$ representing the query result, where N is the size of the database and subw the submodular width of the query. If the arity of all relations is bounded by a constant, then subw is linear in the treewidth tw of the query. In this setting, the authors of this paper proved a lower bound of $N^{\Omega(tw^{\varepsilon})}$ on the circuit size (ICALP 2023), where $\varepsilon>0$ depends on the excluded grid theorem.
Our first main contribution is to improve this lower bound to $N^{\Omega(tw)}$, which is tight up to a constant factor in the exponent. Our second contribution is a $N^{\Omega(subw^{1/4})}$ lower bound on the circuit size for join queries over relations of unbounded arity. Both lower bounds are unconditional lower bounds on the circuit size for well-chosen database instances. Their proofs use a combination of structural (hyper)graph theory with communication complexity in a simple yet novel way. While the second lower bound is asymptotically equivalent to Marx's conditional bound on the decision complexity (JACM 2013), our $N^{\Theta(tw)}$ bound in the bounded-arity setting is tight, while the best conditional bound on the decision complexity is $N^{\Omega(tw/\log tw)}$. Note that, removing this logarithmic factor in the decision setting is a major open problem.